Item

Abstract

All the genes of an organism's genome build up an intricate network of connections between them. Many of these connections are unknown, but knowing about the structure of the network is important for e.g. medical applications. This leads to the problem of reverse engineering the (large-scale) gene regulatory network from gene expression data. Gene network reconstruction can be formulated as a problem of feature selection in a linear regression framework, and we include additional information (like co-binding of transcription factors) about the network with a grouping of features. Available methods for feature selection in the presence of grouping information have different short-comings: Lasso methods underestimate the regression coefficients and do not make good use of the grouping information, and Bayesian approaches often rely on the stochastic and slow Gibbs sampling procedure to recover the parameters, which makes them infeasible for gene network reconstruction.

Here we present a Bayesian method for feature selection with grouping information (with sparsity on the between- and within group level), where the parameters are recovered by a deterministic algorithm (expectation propagation). This sparse-group framework is applied to (large-scale) gene network reconstruction from gene expression data and extended to the vector autoregressive model for time series data.

We prove (on simulated and experimental data) that the Bayesian approach is the best choice for network reconstruction for three reasons: Highest number of correctly selected features, best prediction on new data and reasonable computing time.

We show that a Bayesian approach to feature selection is superior to lasso methods on time series data. Results on experimental temporal data are inconclusive for the linear model.

Finally we note that the presented method is very fundamental and not restricted to the reconstruction of gene regulatory networks, but can be applied to any feature selection problem with grouped features.